Metis

▎ Metis is a general-purpose AI agent built on one thesis: most tasks don't need a frontier model, and the ones that do should be identified before you pay for them — not after. ▎ ▎ Every task first runs on a small quantized model (Qwen2.5-3B) baked into the container, executing locally at zero API cost. Metis samples the local model several times in parallel and measures agreement: consistent answers across independent samples are strong evidence of correctness. That evidence — consensus, token-level confidence, and per-category signals — feeds an isotonic-calibrated confidence gate fitted on real observations. Only when the gate predicts the local answer would miss the accuracy bar does Metis escalate to a paid Fireworks model, chosen per task category with reasoning effort tuned so escalations cost single-digit tokens where possible. ▎ ▎ Known-hard categories (multi-step math, entity extraction) are force-escalated based on measured failure data, math gets a program-aided verification pass first, and an adaptive governor watches the runtime budget and downshifts sampling if the clock runs short. The whole system degrades gracefully: any failure at any layer falls back to the next-best answer rather than crashing. ▎ In our full 48-task rehearsal, Metis hit 95.8% judged accuracy while escalating only 29% of tasks — spending just ~1,400 total paid tokens, versus tens of thousands for a naive route-everything-to-the-cloud agent at comparable accuracy.